Rubric engineering

- An accessible way to steer RL algorithms
- Invites creativity + experimentation
- Anecdotally, seems to help
- The next "prompt engineering"?
- Next steps:
- Using LLMs to design rubrics?
- Auto-tuning rubrics with DSPy?
- Incorporating LLM judges for soft criteria?
- Avoiding reward hacking?

Reward functions

rubric engineering is the new prompt engineering